Idea Grooming

Idea Grooming

You evaluate captured ideas against a rubric and write the analysis back into the source file in place. The original capture stays at the top, untouched. Frontmatter flips from status: raw to status: groomed and gets the structured fields populated.

This skill is domain-agnostic — the rubric works for financial newsletters, perfumery products, fitness apps, AI tools, anything. Domain customization (which questions to emphasize, what counts as novelty in this space, what comparables to surface) lives in the agent's config, not in this skill. Read your config.json or domain-specific notes for that overlay.

Inputs

A markdown file at a path the operator provides (typically <vault>/Ideas/<idea>.md). The file follows the canonical template:

---
title: <one-line idea title>
status: raw
created: YYYY-MM-DD
groomed:
captured_by: <sam | quint | other>
tags: []
originality_score:
market_size:
manual_burden:
verdict:
folder_promoted: false
---

# <Title>

## Original capture
<verbatim source>

## Originality
*(empty — you fill)*

## Market potential
*(empty — you fill)*

## Moats
## Barriers
## Scaling vulnerabilities
## Manual burden
## Verdict

If a file you're asked to groom doesn't have the frontmatter shell, add it. Don't refuse the work just because the format is off.

Posture (read this every time)

• Ground every claim in something concrete. "There's clearly demand" is not a claim. "Stratechery has 30k+ paid subs at $120/yr in this exact category" is.
• Default to skepticism. Most ideas are average. Originality is rare; mark originality_score: 2 if that's the truth.
• Surface disagreement. If two parts of the analysis tension (huge market BUT massive moats already in place), say so.
• No flattery. If verdict is kill, write the kill reason cleanly. The whole point of grooming is sharper thinking, not validation.
• Write the body in plain language. This is read by humans, not by other agents. No bureaucratese.

Section-by-section rubric

Originality (1-5)

• 1 — exists at scale, undifferentiated copy
• 2 — exists, this version has a small twist
• 3 — meaningful differentiation in a crowded space
• 4 — genuinely new angle in a known category
• 5 — new category, no obvious comparable

Always name the closest 2-3 comparables explicitly. If there are none, that's a flag — most "original" ideas just have unfindable competition, not zero.

Market potential

Three reads:

• Who buys/uses — name the persona, not "people who want X"
• Demand signal — what's already paying for nearby goods? Search trends? Subreddit activity? Conference floor density?
• Size band — small (<$10M TAM, lifestyle), mid ($10M–$1B, real business), large ($1B–$50B, venture-scale), massive ($50B+, generational)

Pick the band you can defend. If you'd cringe at a partner reading your defense, drop a band.

Moats

For each defensible advantage, mark it:

• Data — proprietary data accumulating with use
• Distribution — owned audience, established channel
• Brand — recognized name worth a premium
• Network — value increases with users
• Switching cost — pain to leave, not just inertia
• Regulatory — license/compliance barrier
• Talent — rare skill concentrated here
• IP — patents, copyright, trade secret

For each: load-bearing (would actually defend in a price war) or weak (sounds defensible, isn't). Don't list a moat if it's weak — say "no real moat" honestly.

Barriers

What would actually stop us from shipping v1?

• Tech — does it exist, is it ours to build, do we have skills
• Capital — burn rate to MVP
• Distribution — how do customers find this
• Regulation — license/compliance to operate
• Talent — who has to be hired
• Customer trust — does the audience believe a 2-person op can deliver

Rank by which would ACTUALLY stop us. Customer trust often beats tech.

Scaling vulnerabilities

What breaks between MVP and 10x / 100x / 1000x:

• Manual-ops bottleneck (does someone have to touch every customer)
• Unit economics (cost curve flat or flipping negative at scale)
• Support burden (who answers tickets at 1k users)
• Compliance scope (regulation tightens at audience size)
• Vendor lock-in (Mailchimp limits, API rate limits, etc.)

For each: at what scale does this bite? Day 1, 100 users, 10k users?

Manual burden

Honest labor estimate:

• low — <2 hr/week to operate at MVP scale
• med — 2–15 hr/week
• high — 15+ hr/week, requires a hire to scale

Be specific about WHERE the burden sits — content production, customer success, ops, content review. AI helps where it's actually generative; not where it's review/judgment.

Verdict

• pursue — passes on all dimensions OR strong enough on 2-3 to justify exploring deeper. Add 1-3 concrete next moves.
• park — promising but blocked by timing, capital, or another priority. Add a "revisit when X" condition.
• kill — fails core dimension (no moat AND no market AND no originality). Add the one-line kill reason.
• needs-more — analysis genuinely can't conclude with available info. Add what to learn.

Promotion rule (single file → folder)

Promote a .md to a folder (<idea-name>/Index.md + Notes/ + References/ + Open Questions.md) when ALL true:

• verdict: pursue
• 3+ distinct work threads (e.g. tech build + brand work + customer dev + legal review)
• The single-file structure is becoming hard to reason about

Set folder_promoted: true in frontmatter when you do this. Move the original file to become Index.md in the new folder.

Domain overlay (read agent config)

Before grooming, check your agent's config for domain-specific signals:

• domains: [<list>] — what spaces this agent works in (financial, perfumery, fitness, ai-communities, etc.)
• comparables_seed: [<list>] — known companies/products in this space the rubric should consider
• market_signals: [<list>] — things that count as demand evidence in this domain
• kill_dimensions: [<list>] — domain-specific dealbreakers (e.g. "regulated advice without a license" for financial)

If your agent has no domain config, run the rubric domain-agnostic and surface in the verdict that no domain context was loaded.

Output

You write the analysis back into the source file in place:

1. Update frontmatter: status: groomed, groomed: <today's date>, fill originality_score, market_size, manual_burden, verdict, populate tags if useful
2. Fill each section body in the rubric order above
3. If verdict is pursue and 3+ work threads emerge, also do the promotion (move file to folder, create subfolders)
4. Don't touch the ## Original capture section — that's the historical record

Confirm completion in the way your agent normally confirms work (Telegram message to operator, log entry, etc.).

Re-grilling (cron pattern)

When run on a schedule against existing status: groomed + verdict: pursue ideas:

• Re-read the file
• Check for fresh news/market shifts in the domain (use the business-news-monitor skill output if available)
• If the original analysis still holds → update groomed date only
• If something material has shifted (new competitor, regulation change, market shrinkage) → append a ## Re-grill <date> section noting the shift + impact on verdict
• If verdict needs to change → update frontmatter + add the re-grill section explaining why

Don't rewrite the original analysis — append. The trail is part of the value.